Added the residual option for GATConv and GATv2Conv.

[Linnore]

As discussed here #7555 (reply in thread), the GAT paper mentioned using skip-connection to reach the reported metrics on PPI dataset. This PR adds the option of residual to enable skip-connection.

Specifically, we consider the following for adding residual:

1. Define the actual_out_channels = heads * out_channels if concat else out_channels
2. If in_channels == actual_out_channels, then residual is just x before the convolution. Otherwise, residual is a linear projection of the input x onto the dimension of actual_out_channels.

For bipartite case where the input is (x_src, x_dst), the above consideration will be applied for x_dst.

[Linnore]

An example gat_ppi.py is added under examples/gat_ppi.py. Running the example with argument --residual and --no-residual can show that the implementation of residual is effective.

Note that for the case in_channels == actual_out_channels described above, there is no fully connected layer added as in examples/ppi.py to get the residual; instead, an identity mapping is used. Though having slower convergence speed, this is a reasonable design and is the same design as DGL's GATConv just for reference.

I am not sure whether the residual is correctly implemented for the bipartite input though. It's probably problematic when the bipartite input is (x_src, x_dst=None). In GATv2Conv, such input should be asserted and should not be allowed, but GATConv seems not having the same checking.

Feel free to comment!

[Linnore]

The Linting check fails due to torch_geometric/config_mixin.py:85: error: "Type[DataclassInstance]" has no attribute "_target_" [attr-defined]. It is not related to this PR.

[zechengz]

In general LGTM. Maybe @rusty1s can take a look

[Linnore]

Question about adding linear transformation when in_channels == total_out_channels